Anda di halaman 1dari 16

1154 IEEE TRANSACTIONS ON SMART GRID, VOL. 3, NO.

3, SEPTEMBER 2012
Fully Distributed State Estimation
for Wide-Area Monitoring Systems
Le Xie, Member, IEEE, Dae-Hyun Choi, Student Member, IEEE, Soummya Kar, Member, IEEE, and
H. Vincent Poor, Fellow, IEEE
AbstractThis paper presents a fully distributed state estimation
algorithm for wide-area monitoring in power systems. Through it-
erative information exchange with designated neighboring control
areas, all the balancing authorities (control areas) can achieve an
unbiased estimate of the entire power systems state. In comparison
with existing hierarchical or distributed state estimation methods,
the novelty of the proposed approach lies in that: 1) the assumption
of local observability of all the control areas is no longer needed;
2) the communication topology can be different than the physical
topology of the power interconnection; and 3) for DC state esti-
mation, no coordinator is required for each local control area to
achieve provable convergence of the entire power systems states to
those of the centralized estimation. The performance of both DC
and ACstate estimation using the proposed algorithmis illustrated
in the IEEE 14-bus and 118-bus systems.
Index TermsDistributed state estimation, power system com-
munication, wide-area monitoring systems.
I. INTRODUCTION
T
HE electric power industry is undergoing profound
changes as our society emphasizes the importance of
a smarter grid in support of sustainable energy utilization.
Technically, enabled by advanced control, communication,
and computation, wide-area monitoring systems (WAMSs) of
the future are likely to involve many more fast information
gathering and processing devices (e.g., phasor measurement
units) [1]. Institutionally, power industry deregulation has led
to the creation of multiple regional transmission organizations
(RTOs) to operate portions of a large interconnected power
system [2]. Both technical and institutional changes suggest the
need for more decentralized estimation and control in wide-area
power system operations [3].
The main objective of this paper is to propose a fully dis-
tributed approach to state estimation in large multi-area power
systems. State estimation is one of the key functions in control
Manuscript received May 15, 2011; revised November 02, 2011 and March
01, 2012; accepted April 21, 2012. Date of publication May 30, 2012; date of
current version August 20, 2012. This work was supported in part by Power Sys-
tems Engineering Research Center (PSERC), in part by Texas Engineering Ex-
periment Station (TEES), and in part by the DTRAunder Grant HDTRA1-07-1-
0037. Paper no. TSG-00178-2011.
L. Xie and D.-H. Choi are with the Department of Electrical and Computer
Engineering, Texas A&M University, College Station TX 77843 USA (e-mail:
Lxie@ece.tamu.edu; cdh8954@neo.tamu.edu).
S. Kar is with the Department of Electrical and Computer Engi-
neering, Carnegie Mellon University, Pittsburgh, PA 15213 USA (e-mail:
soummyak@andrew.cmu.edu).
H. V. Poor is with the Department of Electrical Engineering, Princeton Uni-
versity, Princeton NJ 08540 USA (e-mail: poor@princeton.edu).
Color versions of one or more of the gures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identier 10.1109/TSG.2012.2197764
Fig. 1. Contrast of communication architectures for hierarchical and fully dis-
tributed state estimation.
centers energy management systems (EMSs) [2], [4]. A state
estimator converts redundant meter readings and other available
information obtained froma supervisory control and data acqui-
sition (SCADA) system into an estimate of the state of an inter-
connected power system [5] and distribution system [6]. While
large power interconnections such as the eastern/western inter-
connections are usually operated by several RTOs, advanced ap-
plications such as wide-area monitoring and control require the
state of the entire system to be available to all the RTOs [3],
[7]. This creates the need for a more decentralized approach to
estimating the entire interconnections state information via ad-
vanced communication. Fig. 1 contrasts the communication ar-
chitecture of fully distributed state estimation with that of hier-
archical state estimation in a multi-area power system.
Several approaches to more decentralized state estimation
(see [8] and [9], for example, for a treatment of decentralized
iterative algorithms for system analysis and optimization) have
been proposed in the literature. In [10] and [11], a star-like
hierarchical state estimation method was proposed. More re-
cently, two-level state estimation for multi-area power system
has been studied in [12][14] and [15], driven by the capability
and need to conduct WAMS. The local state estimation obtained
at the rst level is coordinated at a higher level via synchro-
nized phasor measurements. A survey of multi-area state es-
timation is given in [16]. Most recently, a multilevel state es-
timator (feeder, substation, transmission system organization,
and regional levels) is described for the purpose of monitoring
large-scale interconnected power systems [17]. However, as the
1949-3053/$31.00 2012 IEEE
XIE et al.: FULLY DISTRIBUTED STATE ESTIMATION FOR WIDE-AREA MONITORING SYSTEMS 1155
number of measurements and sampling rate increase, hierar-
chical state estimation approaches may suffer from communi-
cation bottleneck and computational reliability issues inherent
in a system architecture with one single coordination center. A
parallel and distributed state estimation system was envisioned
in [18]. By leveraging the naturally decoupled systemcharacter-
istics of weighted least squares (WLS) estimation, the state esti-
mation problem is decomposed into each areas local estimator
with a coupling constraint optimization technique to ensure con-
vergence of the boundary buses estimates. Numerical results il-
lustrate that the distributed algorithm can not only speed up the
computational time, but also yields acceptable accuracy. How-
ever, local observability of each control area is always required
in the aforementioned algorithms. In other words, all the local
control areas need to have enough measurement redundancy in
order to compute the locally decoupled weighted least squares
estimates (excluding the boundary bus measurements). This as-
sumption may not always hold due to 1) the increasing vul-
nerability of measurements subject to potential bad/malicious
data, and 2) the emergence of smaller control areas such as
micro-grids. A fully distributed algorithm for estimating power
systemdynamic states is proposed in [19] without a requirement
of local observability. However no analytical study has been
conducted for provable convergence of the distributed state es-
timation algorithms to the centralized estimates.
In this paper, a fully distributed static state estimation algo-
rithmwith relaxed local observability is exploited. Starting from
one of the authors recent work [20], an iterative distributed state
estimation scheme is proposed, under which the local control
areas begin with their own estimates of the entire system, com-
municate their estimates with pre-specied neighboring control
areas, and eventually making all local estimates converge to the
centralized state estimation result. In summary, the main contri-
bution of this paper is threefold:
Adistributed, fast state estimation algorithmis proposed. It
does not require either local observability or a central coor-
dinator. As long as the entire interconnection is observable
and the communication graph is connected (which does not
necessarily need to overlap with the physical topology of
the power network), all local areas estimates of the entire
system will converge to the centralized estimates.
The convergence rate of the proposed distributed estima-
tion algorithm is analytically shown for DC weighted least
squares estimation.
The proposed algorithm is implemented for both AC and
DC state estimation for wide-area power systems.
This paper is organized as follows. In Section II, the problem
formulation for distributed state estimation is introduced. The
proposed distributed state estimation algorithm is presented in
Section III. Analytical results on the convergence of the pro-
posed algorithm are also discussed. Illustrative case studies of
both distributed DC and AC state estimation using the proposed
algorithm are presented in Section IV. Concluding remarks and
a discussion of future work are included in Section V.
II. PROBLEM FORMULATION
A. Preliminaries
In -dimensional Euclidean space , the identity ma-
trix is denoted by , while and represent the column vec-
tors with all ones and all zeros in , respectively. The operator
applied to vectors and matrices corresponds to the stan-
dard Euclidean 2-norm and the induced 2-norm, respectively.
The induced 2-norm is equivalent to the matrix spectral radius
for symmetric matrices.
In this paper, we assume that all the random variables are
dened on a common measurable space, . In addition,
all inequalities involving random variables are to be considered
a.s. (almost surely); see [21].
In the spectral graph theory literature, for an undirected graph
is the set of nodes or vertices with
, and is the set of edges with , where
denotes the cardinality. The unordered pair belongs to the
set if nodes and are connected to each other through an
edge. We consider only simple graphs that contain no self-loops
and multiple edges. A graph is connected if there exists a path,
1
between each pair of nodes. The neighborhood of node is
dened as
(1)
Node has degree , the number of neighboring
edges of node . The structure of the graph can be expressed
by the symmetric adjacency matrix , in
which element , if , and , other-
wise. Assuming that the degree matrix is the diagonal matrix
, the graph Laplacian matrix, , is
(2)
Due to a positive semidenite property of the Laplacian matrix,
its eigenvalues can be ordered as
(3)
The smallest eigenvalue is always equal to zero, with
being the corresponding normalized eigenvector.
The multiplicity of the zero eigenvalue equals the number of
connected components of the network; for a connected graph,
. This second eigenvalue is the algebraic connectivity
or the Fiedler value of the network; see [22], [23], or [24] for
detailed treatment of graphs and their spectral theory. For the
computation of vectors, Kronecker products will be involved in
most of the matrix manipulations. For example, the Kronecker
product
2
of the matrix and will be an
matrix, denoted by .
B. Multi-Area Power System State Estimation
An interconnected multi-area power system is assumed to
be partitioned into a total of regions, each region cor-
responding to a geographically non-overlapping control area.
Each control area is allowed, if necessary, to exchange infor-
mation with its neighboring areas. The measurement model for
multi-area state estimation is formulated as follows:
(4)
where is measurement vector (including the boundary injec-
tion and ow measurements) in control area is the state
1
A path between nodes and of length is a sequence
of vertices, such that .
2
The Kronecker product of an matrix with an matrix
is an matrix denoted by whose th block is the
matrix , where denotes the th (scalar) entry of .
1156 IEEE TRANSACTIONS ON SMART GRID, VOL. 3, NO. 3, SEPTEMBER 2012
vector of the entire interconnected power system, is a
nonlinear measurement function for control area , and is a
measurement error vector with zero mean in area .
Initially, we study the linearized DC state estimation problem
with 1.0 per unit (p.u.) voltage magnitudes at all buses and j1.0
p.u. branch impedance. Then, the state vector is taken as the
voltage phase angle vector for all control areas. Therefore, the
nonlinear measurement model for multi-area state estimation
(4) is modied to
(5)
Centralized state estimation computes the optimal estimate
of by minimizing the weighted least squares of measurement
error:
(6)
(7)
where and
. is the positive denite covariance ma-
trix of the noise vector for area . Dening
and for each control area , the centralized
weighted least squares estimate of is given by
(8)
where and . The solution
in (8) is obtained under the following assumption:
Assumption (E.0)Global Observability: The matrix
(9)
is full-rank.
Remark 1: Under assumptions (E.0), the weighted Gramian
(10)
is also full-rank.
It is obvious that the centralized computation for the optimal
estimate requires the knowledge of all the measurement Ja-
cobian matrices , the covariances , and observations
at the central control center. In the next section, we study the
proposed distributed - (Modied-Coordinated State Es-
timation) algorithm, where by inter-control area data exchange,
each control area with only its local measurements and the cor-
responding Jacobian matrix in the network is able to construct
the estimate .
III. DISTRIBUTED ALGORITHM
In this section, we investigate the distributed estimation al-
gorithm - for wide-area monitoring systems. The pri-
mary goal of this section is to showthat for linearized DCpower
ow-based measurement models, it is possible to design totally
distributed iterative schemes where each control area converges
almost surely
3
(a.s.) to the centralized least squares estimator
of the state . In particular, we show that the convergence of
the - algorithm at each control area holds true under
the assumption of global observability and connectivity of the
inter-control area communication network. In [20], [25], and
[26], it is shown that the algorithm - can be extended 1)
to yield more general centralized estimates at each control area,
for example, the maximum likelihood estimate; 2) to deal with
nonlinear observation models; 3) to operate in unpredictable
environments with random inter-control area communication
failure or transmission noise; and 4) to process observations ar-
riving sequentially over time. The second statement mentioned
above implies that the algorithm - can be applicable to
nonlinear AC state estimation. However, due to including the
inversion of the nonlinear measurement functions in a
distributed iterative algorithm, it is intractable to implement dis-
tributed AC state estimation. Instead, we adopt a new approach
of updating a linearized Jacobian matrix without the inversion
of the corresponding nonlinear measurement function. In other
words, each local linearized Jacobian matrix is updated once
every few iterations to assure the convergence of distributed
state estimation to centralized state estimation. In the next sec-
tion, it is shown by simulation results that the performance of
distributed AC state estimation based on the proposed approach
is satisfactory.
A. DC State Estimation
For simplicity, we assume that the vector of initial estimates
of the states, , is deterministic where is the total
number of buses. Asequence of estimate vectors, , is
computed by each control area in a distributed iterative manner.
The state estimate vector of the th control area at the
th iteration is a function of its previous estimate vector,
the communicated estimate vectors at the th iteration from its
neighboring control areas, and the local measurement vector .
Algorithm - : Based on the current state vector ,
the exchanged data , and the measurement vector
, we update the estimate of the states at the th control area
via the following distributed iterative algorithm:
(11)
In (11), are appropriately chosen time-varying
weight sequences (here, time implies iteration). Algorithm (11)
is distributed because for the th control area, it involves only
the data from the sensors in its neighborhood .
3
Note here that all estimates (centralized or distributed) are random objects,
being a functional of the random observations . Hence, any meaningful con-
vergence of such estimate sequences needs to be interpreted in a probabilistic
sense. In this paper, all convergence results are shown to hold in the almost sure
(with probability one) sense, which means convergence for a set of sample paths
or instantiations having probability one.
XIE et al.: FULLY DISTRIBUTED STATE ESTIMATION FOR WIDE-AREA MONITORING SYSTEMS 1157
The iterations in (11) can be written in compact form as
(12)
Here, is the vector of sensor states
(estimates.) The Laplacian matrix captures the topology of
the sensor network. We also dene the matrices and as
(13)
and
(14)
We refer to the recursive estimation algorithm in (12) as
- (Modied-Coordinated State Estimation). We note
that the estimate vector sequence is random, due to the
stochasticity of . Hence, all convergence results will be proved
in an almost sure (a.s.) sense. Based on (11), the procedure of
the proposed distributed state estimation can be summarized as
follows.
Step 1) Each control area knows only its own local Jaco-
bian matrix , local measurement vector , local
covariance matrix , and time varying weight
parameters , and . The aforementioned
data will be kept at every iteration of the proposed
algorithm.
Step 2) A global observability test is conducted in a dis-
tributed manner as follows. Starting from an arbi-
trary positive semi-denite local weighted Gramian
, each local area participates in the following
update:
Then, each area obtains the normalized weighted
Gramian and
computes the rank of to check global observ-
ability. If the network is not globally observable,
observability restoration can be performed using
the method proposed in [27].
Step 3) At the 0th iteration, each area sets the initial esti-
mate vector . We note that this estimate vector
for each control area is partitioned into control
area subvectors conformally with the sets of buses
associated with the control areas. For example,
in Fig. 2, the initial estimate vector corresponding
to area can be expressed as
(15)
Fig. 2. IEEE 14-bus system.
where
and denote area and the initial voltage
phase angle at bus in the th area, respectively.
Then, each area concurrently sends its estimate
vector to the neighboring areas.
Step 4) At the rst iteration, two tasks are sequentially con-
ducted using (11):
a) Computation: each area computes the es-
timate vector based on its previous
estimate vector and the communicated
estimate vectors together with
, and .
b) Communication: each area again sends its
estimate vector to the neighboring areas
for the next iteration process.
Step 5) The distributed iterations in Step 4) are repeated for
a nite number of times, say , such that the con-
trol area estimates approach the centralized least
squares estimate to within a desired level of accu-
racy. Note that the number depends on several
factors including the size of the physical network,
the sparseness of the communication graph, and the
level of accuracy desired. In practice, reasonable ap-
proximations for may be obtained through ofine
training or simulations.
The following assumption on the connectivity of the inter-
area communication network is assumed:
Assumption (E.1)Connectivity: The inter-area communi-
cation network is connected, i.e., .
1158 IEEE TRANSACTIONS ON SMART GRID, VOL. 3, NO. 3, SEPTEMBER 2012
Further, the time-varying weight sequences and
associated to the agreement (consensus) and innovation
potentials,
4
respectively, are assumed to satisfy:
Assumption (E.2)Time Varying Weights: The sequences
and are of the form
(16)
where are constants and the exponents satisfy
(17)
Under (E.0)(E.2), the convergence of DC state estimation is
asserted in the following theorem:
Theorem 2: Consider the - under (E.0)(E.2). Fur-
ther, if , assume, in addition
(18)
where denotes the largest eigenvalue of . Then, for
each , the estimate sequence converges a.s. to the cen-
tralized least squares estimator , i.e.,
(19)
Remark 3: The proof of Theorem 2 is provided in
Appendix A for (note that the case can be
treated similarly, the proof being omitted to avoid unnecessary
repetition). We remark on the consequence of Theorem 2 and
the assumptions . First, we point out the necessity
of the time-varying weight sequences and asso-
ciated to the innovation and consensus potentials, respectively.
In fact, a constant weight version of - , the was
proposed and analyzed in [28]. The state update rule (in vector
form) in such a case reduces to
(20)
where are constants. The following was established
for the in [28]:
Theorem 4 (Corollary 4 in [28]): Let (E.0)(E.1) hold and
the constant satisfy
(21)
where denotes the matrix . Then, if in
addition, the centralized least squares satises , the
estimate sequence for each control area converges to
4
Broadly speaking, the agreement potential corresponds to the agreement
term in (12) and quanties the tendency of the sensor esti-
mates to mutually agree, thus highlighting the collaborative nature of the es-
timation procedure. Innovation potential here corresponds to the effect of ob-
servation data on the estimate update process and is manifested by the quantity
in (12).
at a geometric rate, i.e., there exists constants and
, such that
(22)
for each .
Thus, without the additional assumption , although
the may be shown to converge, the limiting value, in gen-
eral, is not the centralized least squares estimate . As shown
in [28], this is primarily due to the fact that, in general
(23)
and hence, is not a xed point of the update process (20).
Only in cases where (for example, is an invertible
square matrix), the limiting value is , otherwise it deviates
from gracefully.
To overcome this, we resort to time-varying weight sequences
in the - scheme, which (as shown by Theorem 2) con-
verges to for all observable . The use of time varying
weight sequences makes the convergence proof (Appendix A)
much more involved and is beyond the purview of xed point
based techniques as used in [28]. Moreover, as shown in the
proof, the design of and in accordance with as-
sumption (E.2) provides the right interplay between the collab-
orative network information ow and the local sensor observa-
tions leading to the desired convergence to .
Theorem 2 establishes the convergence of the - . As
long as the weight sequences satisfy (E.2), convergence is guar-
anteed. However, the convergence rate depends on the partic-
ular choices of the algorithm parameters, , and . Con-
vergence rate analysis of such mixed time scale procedures is
rather technical and will be addressed in future work. The ex-
tensive numerical simulations presented in the paper show that
the convergence rate is reasonable and also provide guidelines
in choosing the various algorithm parameters.
We comment in this context that under more specic assump-
tions, for example, , the convergence rate can be sped
up further by employing the xed weight version of the
- . In fact, as shown by Theorem 4, the control area esti-
mates are guaranteed to converge at a geometric rate to pro-
vided the condition holds. As pointed out before, in
general, the limiting value, say , of the deviates from
. However, as our simulations demonstrate, as long as
is close to (or, in other words, there is not much redundancy
in the measurement process), the limiting value of the
does not deviate much from . This, in general, points to
the following interesting trade-off between accuracy and con-
vergence rate: for generic measurement models, one may use
the - (with decaying weights) if estimate convergence is
desired to the exact centralized least squares ; on the other
hand, if one is willing to tolerate a little error (small deviation
from ), one may instead use the which exhibits geo-
metric (exponential) convergence.
XIE et al.: FULLY DISTRIBUTED STATE ESTIMATION FOR WIDE-AREA MONITORING SYSTEMS 1159
B. AC State Estimation
Alternatively, for AC state estimation, the iterative equation
(11) is modied to
(24)
Here, is the transformed Jaco-
bian matrix of . In addition,
= is the time-varying coefcient where
implies modulo , and is a xed number
of iterations. For each interval
is updated as follows:
(25)
where represents the th xed iteration interval.
C. Bad Data Processing
In the presence of bad data, the proposed distributed frame-
work could also detect and identify bad data in a similar manner.
Due to space limitations, we do not elaborate on this issue here.
Interested readers are referred to [29][31] for detailed analysis
of this situation.
IV. CASE STUDIES
In this section, we analyze the performance of the proposed
distributed state estimation in the IEEE 14-bus and IEEE
118-bus systems. The centralized WLS state estimate pro-
vides a performance benchmark for the proposed algorithm.
In Section IV-A, we introduce performance indices used for
assessing the performance of the proposed distributed state es-
timation algorithm. The performance of the proposed algorithm
is analyzed in the two subsequent subsections, corresponding
to DC state estimation and AC state estimation, respectively.
The rst part of these two subsections includes the results of
the observability analysis for the IEEE 14-bus and 118-bus
systems. This is followed by detailed numerical simulations to
investigate the convergence rate of the proposed algorithm.
A. Performance Evaluation
The performance of the proposed distributed state estimation
algorithm is evaluated in terms of the following performance
indices:
1) Estimation Accuracy: We choose the bus phase angle and
voltage magnitude difference between distributed and central-
ized algorithms as the performance indices to evaluate the con-
vergence of the proposed algorithm:
(26)
where subscripts and correspond to buses and , respec-
tively. and
represent the absolute values of bus and bus s phase angle
Fig. 3. Multi-area IEEE 118-bus system illustrating two different inter-control
communication networks.
differences in distributed and centralized state estimation, re-
spectively, and
(27)
where and represent bus s estimated voltage
magnitudes in distributed and centralized state estimation,
respectively.
2) Execution Time Efciency: The execution time efciency
of the proposed algorithm is dened as
% (28)
where and represent the system-wide execution time of
the centralized and distributed state estimation algorithms, re-
spectively. Note that the system-wide execution time of the dis-
tributed state estimation algorithm is always equal to the total
execution time of the slowest area among local control areas.
B. Case 1: DC State Estimation
In this case, we assume a DC state estimation model with
1.0 p.u. voltage magnitudes in all buses and p.u. branch
impedance. For the IEEE 14-bus system, the measurement con-
guration including the types as well as locations of the mea-
surements and network decompositions comes from [12]. The
systemhas four nonoverlapping control areas as shown in Fig. 2.
Control areas 1, 2, 3, and 4 contain ,
and buses, respectively. On the other hand, the IEEE
118-bus system has nine nonoverlapping control areas, as con-
sidered in [32] and shown in Fig. 3. The control areas contain
, and buses, respectively. Initially, mea-
surement noises are assumed to be Gaussian with zero mean
and uniform nite variances . We assume that one or mul-
tiple local control areas become locally unobservable for both
systems, but that the entire system is still globally observable.
1) IEEE 14-Bus System: We conduct global and local ob-
servability analysis in the IEEE 14-bus system. We dene the
1160 IEEE TRANSACTIONS ON SMART GRID, VOL. 3, NO. 3, SEPTEMBER 2012
Fig. 4. Multi-area IEEE 14-bus system illustrating two different inter-control
communication networks.
local observability based on [12, (24)]. An area is observ-
able if and only if
(29)
where is the local Jacobian matrix related to all the internal
measurements of area (excluding boundary bus measure-
ments). This system has a total of 22 measurements, including
6 power injection and 16 power ow measurements. Then, the
rank of the system-wide measurement Jacobian matrix is 13
so that the system is still globally observable. The rank test re-
sult for each local measurement Jacobian matrix is shown
as follows:
;
;
;
.
Therefore, area shaded in Fig. 2 becomes locally
unobservable.
Next, we show in this relaxed observability setup that by the
proposed iterative algorithm, all the control areas estimates of
the system-wide state will converge to the centralized WLS so-
lution. We assume that all the power ow and injection mea-
surements are corrupted by additive Gaussian noises with equal
variances . The IEEE 14-bus system has a total
of 22 measurements, which is comprised of 16 branch ow
and 6 nodal power injection measurements. The detailed bus
number and measurement type of each area are shown in Fig. 4.
Constants , and are chosen for the distributed itera-
tive algorithm with , and ,
respectively. In addition, the tolerance of simulation is set to
. For both centralized and distributed state estima-
tion, bus 1 is selected as the slack bus.
The performance test is conducted for a total of 91 pairs of
phase angle differences . We ran-
domly illustrate ve pairs , and in the
gure. In Fig. 5, it is observed that after a short period of os-
cillation (30 iterations), the distributed estimates exponentially
Fig. 5. Convergence of bus phase angle difference between distributed and cen-
tralized algorithms in the IEEE-14 bus system of Case 1.
converge to the centralized estimates. Furthermore, in our re-
cent work [28] (based on the algorithm), we investigated
the sensitivity of the convergence rate of the proposed algorithm
with respect to 1) step size coefcients and ; 2) the measure-
ment error covariance matrix ; and 3) different communication
topologies. The main results for the aforementioned sensitivity
analysis are summarized as follows:
Sensitivity of convergence rate with respect to a and b:
Larger values of and [provided they satisfy (21)] lead to
higher convergence rate at the expense of potentially more
oscillations, and vice versa.
Sensitivity of the convergence rate with respect to values
in the covariance matrix: The convergence rate seems to
be robust to the measurement error covariance matrix.
Convergence rate of the algorithmwith different communi-
cation topologies: The communication topology is elastic
for the convergence of the proposed distributed state esti-
mates to the centralized state estimates on condition that
the inter-area power network is connected, and the whole
system is globally observable.
2) IEEE 118-Bus System: For the IEEE 118-bus system, we
assume that power injection measurements are placed at all gen-
erator buses, power ow measurements at a subset of transmis-
sion lines. Therefore, this system has a total of 178 measure-
ments, including 49 power injection and 129 power ow mea-
surements. Based on this measurement conguration, the rank
of the system-wide measurement Jacobian matrix is 117 so
that the system is globally observable; however, shaded areas
, and in Fig. 3 are identied to be locally un-
observable by the following rank test:
;
;
;
;
;
;
;
;
.
XIE et al.: FULLY DISTRIBUTED STATE ESTIMATION FOR WIDE-AREA MONITORING SYSTEMS 1161
TABLE I
PERFORMANCE IN THE IEEE 14-BUS SYSTEM OF CASE 1
TABLE II
PERFORMANCE IN THE IEEE 118-BUS SYSTEM OF CASE 1
Fig. 6. Convergence of bus phase angle differences between distributed and
centralized algorithms in the IEEE-118 bus system of Case 1.
In the above observability setup, we again examine the con-
vergence rate of the proposed distributed state estimation algo-
rithm with centralized WLS state estimation. Constants ,
and in the proposed distributed iterative scheme are chosen
with , and , and the tolerance of
the simulation is set to . Five pairs ,
and are illustrated in the gure. Fig. 6 shows that in the
IEEE 118-bus system, the distributed estimates also converge
well (in 30 iterations) to the centralized estimates.
Tables I and II summarize the performance of the proposed
distributed state estimation algorithm in the DC state estima-
tion model, corresponding to the IEEE 14-bus and 118-bus sys-
tems, respectively. In these tables, three performance indices
are used: 1) maximum phase angle difference between the true
state values and the distributed state estimates ; 2) average
phase angle difference between the true state values and the dis-
tributed state estimates ; and 3) execution time efciency
dened in (28). We can see from these tables that increasing the
parameters and leads to improved efciency. In addition, it is
observed that the estimation accuracy of the proposed algorithm
degrades if area or boundary measurements have a higher
noise variance . In particular, these tables show the per-
formance of the proposed algorithm for two different commu-
nication topologies where communication scheme 1 graph is
denser than communication scheme 2 graph. However, commu-
nication scheme 1 does not always outperform communication
scheme 2. It can also be observed by comparing the last rows
of Tables I and II, that the efciency gain is more signicant
in the IEEE 118-bus system than in the IEEE 14-bus system.
Therefore, we conjecture that the proposed algorithm is more
suitable for large-scale power networks to achieve the improved
efciency.
C. Case 2: AC State Estimation
In this subsection, we study the performance of AC state es-
timation using the proposed algorithm. First, we consider the
IEEE 14-bus system for the performance analysis of the pro-
posed algorithm. This system is assumed to have the same mea-
surement conguration as the IEEE 14-bus system illustrated
in Case 1 except assuming that power injection and ow mea-
surements are always in pairs and one voltage magnitude mea-
surement is placed at each local control area. Suppose that a
pair of real and reactive ow measurements, and in
area are deleted, which leads to a locally unobservable area
, but a globally observable network. This system has a total
of 48 measurements, including 6 pairs of power injection, 16
pairs of power ow, and 4 voltage magnitude measurements.
Similarly, based on the measurement conguration of the IEEE
118-bus system illustrated in Case 1, the IEEE 118-bus system
in Case 2 has a total of 365 measurements, including 49 pairs of
power injection, 129 pairs of power ow, and 9 voltage magni-
tude measurements.
For distributed nonlinear AC state estimation, we propose
two types of update rules with respect to the local Jacobian
matrix for area , corre-
sponding to: 1) wait-and-update (WAU); and 2) no-wait-and-
update (NWAU) rules. In the WAU rule-based algorithm, each
local control area does not update its own local Jacobian matrix
until after a xed number of iterations. For example, suppose
that this xed number of iterations is equal to 40. In the WAU
rule-based algorithm, every local Jacobian matrix remains un-
changed while the WAU rule-based algorithm counts the itera-
tion from 1 to 39. When the number of iterations arrives at 40,
the local matrix is nally updated with its local state estimate
1162 IEEE TRANSACTIONS ON SMART GRID, VOL. 3, NO. 3, SEPTEMBER 2012
Fig. 7. Convergence of bus phase angle differences between distributed and
centralized algorithms in the IEEE 14-bus system of Case 2.
vector computed at the 39th iteration. Then, the count of itera-
tions resets to zero and the proposed algorithm restarts from the
rst iteration. This updating process continues until this algo-
rithm converges. Here, the xed iteration period is denoted by
FP. On the other hand, in the NWAU rule-based algorithm, each
local control area updates its own local Jacobian matrix at every
iteration. The performance of the WAU and NWAU rule-based
algorithms is analyzed in the next two subsections.
1) IEEE 14-Bus System: Considering both bus voltage mag-
nitude and phase angle as state variables, an area is locally
observable if and only if
(30)
where is the local Jacobian matrix
associated with only all the internal measurements of area
and with at start.
The rank of the system-wide measurement Jacobian matrix
is 27 so that the system is globally ob-
servable. All the local Jacobian matrices satisfy condition (30)
except area as follows:
;
;
;
.
Therefore, area is locally unobservable.
Next, the performance of the WAU and NWAU rule-based
algorithms is evaluated and compared with each other. For the
NWAU rule-based algorithm, it is hard to nd optimum ini-
tial state and parameters and . However, we already know
from the results of the previous subsection that the proposed
distributed state estimation algorithm based on a linearized DC
power ow model converges to a centralized WLS state estima-
tion solution. This fact is our motivation for adopting the WAU
rule-based algorithm rather than the NWAU rule-based algo-
rithm. For illustrating the convergence of the WAU rule-based
algorithm in the gure, ve pairs , and
are randomly chosen. The parameters of the proposed algorithm
are chosen with , and
Fig. 8. Convergence of with varying FP in the IEEE 14-bus system of
Case 2.
Fig. 9. Convergence of and in the IEEE 14-bus system illustrating
the two different communication schemes.
. Fig. 7 shows the convergence of the distributed esti-
mation algorithmto the centralized estimate. FromFig. 8, which
shows the effect of varying FP, two noticeable phenomena are
observed. First, the convergence rate of is changing as the
value of FP is increasing. In other words, the convergence rate
becomes slower as a larger value of FP is chosen. From this
gure, the best value of FP among those considered for the
convergence rate is 4. Second, the WAU rule-based algorithm
outperforms the NWAU rule-based algorithm in terms of phase
angle accuracy. The NWAU rule-based algorithm (without FP)
shows a poor performance as shown in Fig. 8. Therefore, for
faster convergence of the proposed algorithm, the values of pa-
rameters and FP must be carefully chosen. Fig. 9 shows
that the WAU rule-based algorithm is robust to change in the
communication scheme. In Fig. 10, we can observe that voltage
magnitude estimates using the proposed algorithm are rather in-
accurate at some buses.
2) IEEE 118-Bus System: In the IEEE 118-bus test system,
the rank of the system-wide measurement Jacobian matrix
XIE et al.: FULLY DISTRIBUTED STATE ESTIMATION FOR WIDE-AREA MONITORING SYSTEMS 1163
Fig. 10. Bus voltage magnitude differences between distributed and central-
ized algorithms in the IEEE 14-bus system of Case 2.
Fig. 11. Convergence of bus phase angle differences between distributed and
centralized algorithms in the IEEE 118-bus system of Case 2.
is 235 so that the system is globally observ-
able. The rank of every local area Jacobian matrix is as follows:
;
;
;
;
;
;
;
;
.
Since areas and have rank deciency,
both areas become locally unobservable. Five pairs
, and are randomly chosen for
the convergence test. The parameters of the WAU rule-based
algorithm are set to , and
. Fig. 11 shows that after some initial oscillation,
the distributed estimation algorithm converges exponentially
to the centralized estimates. Unlike Fig. 6, this gure shows
Fig. 12. Convergence of with varying FP in the IEEE 118-bus system of
Case 2.
Fig. 13. Convergence of and in the IEEE 118-bus system illus-
trating the two different communication schemes.
some spikes in the plot after every 20 iterations. This is
due to the fact that each local Jacobian matrix is updated at
every 20 iterations. Fig. 12 shows the impact of FP on the
convergence of the distributed state estimation algorithm. In
this gure, we conclude that is the best among those
considered as well as the lower bound for the convergence
of the proposed algorithm since the proposed algorithm
diverges when and the increase of FP leads to slower
convergence rate. Similar to Fig. 9, Fig. 13 shows that the
proposed algorithm is robust to changes in the communication
scheme. As shown in Fig. 14, the WAU rule-based algorithm
in the IEEE 118-bus system also provides a poor voltage
magnitude estimate at some buses.
Tables III and IVsummarize the performance of the proposed
distributed state estimation algorithm in the AC state estima-
tion model, corresponding to the IEEE 14-bus and 118-bus sys-
tems, respectively. Compared to Tables I and II, two extra per-
formance indices are added to the last two rows of Tables III
and IV: 1) maximum voltage magnitude difference between the
1164 IEEE TRANSACTIONS ON SMART GRID, VOL. 3, NO. 3, SEPTEMBER 2012
TABLE III
PERFORMANCE IN THE IEEE 14-BUS SYSTEM OF CASE 2
TABLE IV
PERFORMANCE IN THE IEEE 118-BUS SYSTEM OF CASE 2
Fig. 14. Bus voltage magnitude differences between distributed and central-
ized algorithms in the IEEE 118-bus system of Case 2.
true state values and distributed state estimates ; and 2)
average voltage magnitude difference between true state values
and distributed state estimates . Table III shows that the
WAU rule-based algorithm provides a better phase angle esti-
mation accuracy than the NWAUrule-based algorithm. Table IV
shows that with the same parameters and FP value, the WAU
rule-based algorithm converges whereas the NWAU rule-based
algorithm diverges. This result is one of the advantages for the
WAU rule-based algorithm. The last observation from these ta-
bles is that we can also see that as the network size increases,
the improvement in efciency becomes more signicant as for
DC state estimation.
Lastly, we investigate the communication requirements of our
proposed algorithm. We emphasize here that the algorithm is
quite efcient in terms of communication, as only estimate vec-
tors are exchanged in each iteration. This is in contrast to other
distributed approaches which require the transmission of the
system matrices from sensor to sensor as
these are available only locally. Further, the update rule does
not require matrix inversions, a key computational bottleneck
even in centralized applications. The communication overhead
usually consists of four processing times: 1) the data processing
time; 2) the queueing delay; 3) the transmission time; and 4)
the propagation time. In this paper, only the transmission time
is assumed to be the communication time. Therefore, the com-
munication time is dened as shown in (31) at the bottom of the
page. We assume that the data (elements of estimate vector) are
expressed as 32-bit real numbers and optical bers (
bps) are chosen as communication channels among local control
areas. Table V shows the number of exchanged data values and
the communication time per one iteration in the IEEE 14-bus
and 118-bus systems for both Case 1 and Case 2. Due to the ex-
change of only estimate vectors, the amount of exchanged data
among local control areas is always equal to the number of states
in the power system. This fact implies that the communication
time in our proposed algorithm is not affected by the network
topology and measurement conguration, unlike [12] in which
the communication time depends on the number of boundary
measurements and decomposed local control areas.
Table VI shows the total execution time, the total communi-
cation time, and the total time (i.e., the sumof the total execution
(31)
XIE et al.: FULLY DISTRIBUTED STATE ESTIMATION FOR WIDE-AREA MONITORING SYSTEMS 1165
TABLE V
COMMUNICATION TIME BETWEEN LOCAL CONTROL AREAS
TABLE VI
TOTAL EXECUTION AND COMMUNICATION TIME IN CASE 1
time and the total communication time) for each DC state esti-
mation scheme in the IEEE 14-bus and 118-bus systems. Here,
we assume that the IEEE 14-bus and 118-bus systems have a
total of 22 and 178 measurements, respectively. Since measure-
ments collected by the SCADA system are transmitted to the
centralized state estimator only once, the total communication
time in the centralized state estimation is simply com-
puted by (31). On the other hand, the total communication time
in the proposed distributed state estimation algorithm is calcu-
lated by multiplying the communication time required for one
iteration by the total number of iterations required for the pro-
posed algorithms convergence. The fourth column of Table VI
shows these total communication times with a total of 1 and 50
iterations, corresponding to the centralized and proposed dis-
tributed state estimation, respectively. From this table, we can
see that for both state estimation schemes, the total communi-
cation time is much smaller than the total execution time so that
the total time is dominated by the total execution time. In view
of the total time, the proposed distributed state estimation al-
gorithm still appears to be more efcient for large-scale power
networks.
Finally, the novelty of the proposed distributed estimation al-
gorithm can be summarized as follows:
Fully distributed algorithmwith lower computational com-
plexity: No central coordinator is required to be present in
order for each local control center to achieve entire system
state estimation. Furthermore, the proposed algorithm is
computationally more efcient compared with centralized
estimation, since it does not require extensive matrix in-
version (e.g., gain matrix inversion). As the network size
increases, the relative computational saving compared with
centralized estimation becomes more signicant.
Flexible communication topology: The proposed dis-
tributed estimation algorithm is applicable for any
communication topology provided that the inter-area
communication graph is connected, and the whole system
is globally observable.
Local observability assumption relaxed: Unlike many
existing distributed or hierarchical state estimation algo-
rithms, in our proposed algorithm, all the control areas
are not required to be locally observable. Therefore, the
potential performance degradation due to pseudo-mea-
surement placement can be prevented.
V. CONCLUDING REMARKS
In this paper, a fully distributed algorithm has been pro-
posed for multi-area state estimation in interconnected power
systems. Compared with existing hierarchical or distributed
state estimation algorithms, our proposed algorithm is imple-
mentable under the more relaxed assumption that not all the
control areas must be simultaneously locally observable. In the
case of DC state estimation, we have proven the convergence
of the proposed algorithm to the centralized state estimation
result. In the case of AC state estimation, we have proposed an
implementable distributed WAU rule-based algorithm which
shows satisfactory convergence behavior compared with cen-
tralized state estimation results. Illustrative examples in both
IEEE 14-bus and 118-bus systems conrm the effectiveness of
the proposed algorithm.
Future work should address the question of designing the
most efcient communication topology for fast and robust con-
vergence of the distributed state estimation algorithm. Also, the
practical implementation of the proposed algorithm should be
tested in large-scale realistic AC state estimation. Last but not
least, we plan to integrate bad data processing with the state es-
timation algorithm in a distributed framework (preliminary re-
sults along these lines are reported in [29], [30], and [31]).
APPENDIX A
ANALYSIS AND PROOFS
This section is devoted to the proof of Theorem 2. Due to the
mixed time-scale behavior, the - does not fall under the
purview of standard techniques (as used for the in [28])
for establishing convergence of iterative schemes. The proof is
lengthy and requires a thorough understanding of the two dif-
ferent potentials, agreement and innovations, and their interac-
tion over different time scales. The proof is accordingly accom-
plished in steps.
We provide a brief outline of the major arguments required to
establish the convergence of the - . The rst step consists
of showing that the sensors reach agreement in the long run, i.e.,
as , the sensors follow an averaged or mean behavior,
the estimate trajectories over different sensors merging towards
each other. This part is accomplished in Section A-2, where it is
shown that under the assumption (E.2), the agreement potential
eventually dominates the local innovation potentials, leading the
estimates to consensus. Once consensus is achieved, the next
step is to show the convergence of the averaged system to the
desired least squares estimate. This is further accomplished (see
Section A-3) in two steps: the recursion satised by the averaged
estimate is quite different from that of a recursive centralized
estimator. In the rst step, we construct a ctitious recursive
centralized estimator, which is shown to converge to . The
convergence of the averaged estimate (and hence, the individual
sensor estimates) to is established by a pathwise comparison
with the aforementioned centralized estimate sequence.
1) Some Intermediate Results: We state two results from[20]
to be used for establishing the convergence of - .
The rst result provides convergence conditions of general
time-varying scalar recursions.
Lemma 5 [20, Lemma 4]: Let the sequences and
be given by
(32)
1166 IEEE TRANSACTIONS ON SMART GRID, VOL. 3, NO. 3, SEPTEMBER 2012
where and . Then, if ,
there exists , such that, for sufciently large non-negative
integers,
(33)
Moreover, the constant can be chosen independently of .
Also, if , then, for arbitrary xed
(34)
(We use the convention that , for .)
The next result presents estimates of ellipticity of a sequence
of relevant time-varying matrices, to be used in the sequel.
Lemma 6 [20, Lemma 6]: Under (E.0)(E.2), there exists
sufciently large and a constant , such that, for
(35)
2) Estimate Consensus: We state the key result of this sub-
section, Lemma 7. To this end, denote by , the se-
quence of network-averaged estimates, i.e.,
(36)
Lemma 7: Under (E.0)(E.2), the sensor estimates achieve
consensus, i.e.,
(37)
Remark 8: Lemma 7 presents the interesting fact, that, irre-
spective of different initial estimates and observations, all the
sensors in the network eventually reach agreement (consensus)
in terms of their estimates of the parameter .
The rest of the subsection is devoted to the proof of Lemma
7. We rst establish the boundedness of the estimate sequence
:
Lemma 9: Under (E.0)(E.2), the sensor estimates are path-
wise bounded, i.e., there exists a random variable , such that
(38)
Proof: Since, the observation noise is nite a.s., there ex-
ists a random variable , such that
(39)
Note that (12) may be written as
(40)
We then have from (39)
(41)
By Lemma 6, there exists sufciently large, such that, for all
(42)
where is a constant.
Also, note that, under (E.2), the weight sequence falls
under the purview of Lemma 5 and, hence, there exists suf-
ciently large, such that, for every
(43)
where is a constant and independent of .
Choose and note that the recursion (41) may
be upper-bounded as
(44)
for . The above leads to (for )
(45)
The second term in (45) is bounded by (43) and the niteness
of . For the rst term, we note that
(46)
where the rst step follows from the identity for
small positive [note that the -s go to zero by (E.2)]. The
last step is a consequence of the persistence condition, namely,
for all .
Hence, from (43), (45), and (46), we nally obtain
(47)
Since the above holds pathwise (for all sample paths), we con-
clude that
(48)
which establishes the claim in (38).
We now complete the proof of Lemma 7, which shows even-
tual agreement between the sensor estimates.
XIE et al.: FULLY DISTRIBUTED STATE ESTIMATION FOR WIDE-AREA MONITORING SYSTEMS 1167
Proof of Lemma 7: Dene the sequence by
(49)
where is dened in (36). Recall the matrix
(50)
and note that
(51)
and
(52)
for all .
From (12), the sequence satises
(53)
where we use the identity
(54)
and corresponds to the zero matrix of appropriate dimension.
Using (51) and (53), the sequence may be shown to
satisfy the following recursion:
(55)
where the residual is of the form
(56)
By Lemma 9, the random objects and are dominated
in norm by a.s. nite random variables and , respectively.
Hence, there exists another a.s. nite random variable , such
that
(57)
From (55) and (57), we conclude
(58)
For sufciently large, using the properties of the matrices
and and the Kronecker product manipulations demon-
strated in Section II-A, we have
(59)
and note that , by the connectivity of the communica-
tion network [(E.1)]. It then follows from (58), that, for
(60)
The above recursion leads to
(61)
Note that the last step follows from the fact that the term
goes to zero as , which is due to
positivity of (due to the network connectivity) and the
persistence condition [(E.2)], namely, .
The remaining term on the right-hand side of (61) falls under
the purview of Lemma 5 (choose , and note
that ) and hence goes to zero also. We thus conclude
from (61), that
(62)
Since (62) holds on sample paths (except on a set of measure
zero), the claim in (37) follows.
3) Averaged Estimate Behavior: Convergence of - :
We start by introducing a ctitious centralized estimator,
and establish its convergence to . This estimator will be used
subsequently to prove convergence of the averaged estimator,
through some comparison techniques.
To this end, dene the estimator sequence, , by
(63)
where is the Gramian introduced in (E.0).
The following holds:
Lemma 10: Under (E.0)(E.2), the sequence con-
verges a.s. to , i.e.,
(64)
Proof: By denition of , the least squares estimate of ,
we have
(65)
Dene the process by
(66)
Using (63) and (65), we have the following recursion for
(67)
1168 IEEE TRANSACTIONS ON SMART GRID, VOL. 3, NO. 3, SEPTEMBER 2012
Since is positive denite (full rank) by (E.0), there exists
sufciently large, such that, for
(68)
where is a constant. From (67), it then follows
(69)
Due to the persistence condition on the weights , i.e., that
they sum to , we can take limits on both sides of (69) and
conclude
(70)
Since the above holds a.s., the claim in (64) follows.
We now show that the averaged estimate sequence merges
with the ctitious centralized estimator constructed in (63).
Lemma 11: Under (E.0)(E.2), we have the following:
(71)
Proof: We note that
(72)
Dene the sequence by
(73)
From (63) and (73), we have
(74)
Note that the a.s. boundedness of the sequences
(Lemma 9) and (Lemma 10) imply the a.s. boundedness
of , i.e., there exists a random variable , such that
(75)
Since is positive denite (full rank) by (E.0), there exists
sufciently large, such that, for
(76)
where is a constant.
Now consider , arbitrarily small. The eventual agree-
ment between the sensor estimates (Lemma 7) shows the ex-
istence of (depending on the sample path), sufciently large
(greater than ), such that, for
(77)
Also, by Lemma 5, there exists sufciently large (greater than
), such that
(78)
for and is a constant independent of . Let
. Then, by (74) and (77), we have for
(79)
The above recursion leads to
(80)
As , the rst term on the right-hand side of (80) goes to
zero [see similar arguments in (69) in Lemma 10]. Hence, we
have
(81)
Since the above holds for all , by taking to 0, we have
(82)
This establishes the claim.
We now complete the proof of Theorem 2, which establishes
the convergence of the - estimates to at every sensor.
Proof of Theorem 2: Note that, by Lemma 7 and Lemma
11, it follows that
(83)
The result is then a consequence of Lemma 10 and (83).
REFERENCES
[1] A. Bose, Smart transmission grid applications and their supporting
infrastructure, IEEE Trans. Smart Grid, vol. 1, no. 1, pp. 1119, Jun.
2010.
[2] F. F. Wu, K. Moslehi, and A. Bose, Power system control centers:
Past, present, and future, Proc. IEEE, vol. 93, no. 11, pp. 18901908,
Nov. 2005.
[3] V. Terzija, G. Valverde, D. Cai, P. Regulski, V. Madani, J. Fitch, S.
Skok, M. M. Begovic, and A. Phadke, Wide-area monitoring, protec-
tion, and control of future electric power networks, Proc. IEEE, vol.
99, no. 1, pp. 8093, Jan. 2011.
[4] F. F. Wu, Power system state estimation: A survey, Int. J. Elect.
Power Energy Syst., vol. 12, no. 2, pp. 8087, Apr. 1990.
[5] F. C. Schweppe, J. Wildes, and D. B. Rom, Power system static state
estimation, Parts I, II and III, IEEE Trans. Power App. Syst., vol.
PAS-89, no. 1, pp. 120135, Jan. 1970.
[6] C. N. Lu, J. H. Teng, and W.-H. E. Liu, Distribution state estimation,
IEEE Trans. Power Syst., vol. 10, no. 1, pp. 229240, Feb. 1995.
[7] A. Bose, A. Abur, K. Y. K. Poon, and R. Emami, Implementation Issues
for Hierarchical State Estimators, PSERC Final Project Report, Aug.
2010.
[8] D. P. Bertsekas and J. N. Tsitsiklis, Parallel and Distributed Compu-
tation: Numerical Methods. Belmont, MA: Athena Scientic, 1997.
[9] Y. Saad, Iterative Methods for Sparse Linear Systems. Philadelphia,
PA: SIAM, 2003.
XIE et al.: FULLY DISTRIBUTED STATE ESTIMATION FOR WIDE-AREA MONITORING SYSTEMS 1169
[10] T. V. Cutsem, J. L. Horward, and M. Ribbens-Pavella, A two-level
static state estimator for electric power systems, IEEE Trans. Power
App. Syst., vol. PAS-100, no. 8, pp. 37223732, Aug. 1981.
[11] T. V. Cutsem and M. Ribbens-Pavella, Critical survey of hierarchical
methods for state estimation of electric power systems, IEEE Trans.
Power App. Syst., vol. PAS-102, no. 10, pp. 247256, Oct. 1983.
[12] G. N. Korres, A distributed multiarea state estimation, IEEE Trans.
Power Syst., vol. 26, no. 1, pp. 7384, Feb. 2011.
[13] T. Yang, H. Sun, and A. Bose, Transition to a two-level linear state
estimatorPart I: Algorithm, IEEE Trans. Power Syst., vol. 26, no.
1, pp. 4653, Feb. 2011.
[14] T. Yang, H. Sun, and A. Bose, Transition to a two-level linear state
estimatorPart II: Algorithm, IEEE Trans. Power Syst., vol. 26, no.
1, pp. 5462, Feb. 2011.
[15] A. Gmez-Expsito and A. de la Villa Jan, Two-level state estimation
with local measurement pre-processing, IEEE Trans. Power Syst., vol.
24, no. 2, pp. 676684, May 2009.
[16] A. Gmez-Expsito, A. de la Villa Jan, C. Gmez-Quiles, P.
Rousseaux, and T. V. Cutsem, A taxonomy of multi-area state
estimation methods, Elect. Power Syst. Res., vol. 81, pp. 10601069,
Apr. 2011.
[17] A. Gmez-Expsito, A. Abur, A. de la Villa Jan, and C. Gmez-
Quiles, A multilevel state estimation paradigm for smart grids, Proc.
IEEE, vol. 99, no. 6, pp. 952976, Jun. 2011.
[18] D. M. Falcao, F. F. Wu, and L. Murphy, Parallel and distributed state
estimation, IEEE Trans. Power Syst.., vol. 10, no. 2, pp. 724730,
May 1995.
[19] U. A. Khan, M. D. Ilic, and J. M. F. Moura, Cooperation for ag-
gregating complex electric power networks to ensure system observ-
ability, in Proc. 1st Int. Conf. Infrastructure Systems, Rotterdam, The
Netherlands, Nov. 2008, pp. 16.
[20] S. Kar and J. M. F. Moura, Convergence rate analysis of distributed
gossip (linear parameter) estimation: Fundamental limits and trade-
offs, IEEE J. Select. Topics Signal Process.: Signal Processing in Gos-
siping Algorithms Design and Applications, vol. 5, no. 4, pp. 674690,
Aug. 2011.
[21] P. Billingsley, Convergence of Probability Measures. New York:
Wiley, 1999.
[22] F. R. K. Chung, Spectral Graph Theory. Providence, RI: Amer.
Math. Soc., 1997.
[23] B. Mohar, , Y. Alavi, G. Chartrand, O. R. Oellermann, and A. J.
Schwenk, Eds., The Laplacian spectrum of graphs, in Graph
Theory, Combinatorics, and Applications. New York: Wiley, 1991,
vol. 2, pp. 871898.
[24] B. Bollobas, Modern Graph Theory. New York: Springer-Verlag,
1998.
[25] S. Kar, J. M. F. Moura, and K. Ramanan, Distributed parameter esti-
mation in sensor networks: Nonlinear observation models and imper-
fect communication, IEEE Trans. Inf. Theory, accepted for publica-
tion.
[26] S. Kar, J. M. F. Moura, and H. V. Poor, Distributed linear parameter
estimation: Asymptotically efcient adaptive strategies, SIAMJ. Con-
trol Optim., submitted for publication.
[27] B. Gou and A. Abur, An improved measurement placement algorithm
for network observability, IEEE Trans. Power Syst., vol. 16, no. 4, pp.
819824, Nov. 2001.
[28] L. Xie, D.-H. Choi, and S. Kar, Cooperative distributed state estima-
tion: Local observability relaxed, in Proc. IEEE Power and Energy
Society General Meeting, Detroit, MI, 2011.
[29] D.-H. Choi and L. Xie, Fully distributed bad data processing for wide
area state estimation, in Proc. 2nd IEEE Int. Conf. Smart Grid Com-
munications, Brussels, Belgium, Oct. 2011.
[30] A. Tajer, S. Kar, H. V. Poor, and S. Cui, Distributed joint cyber attack
detection and state recovery in smart grids, in Proc. 2nd IEEE Int.
Conf. Smart Grid Communications, Brussels, Belgium, Oct. 2011.
[31] L. Xie, D.-H. Choi, S. Kar, and H. V. Poor, Bad/malicious data de-
tection in distributed power system state estimation, in Smart Grid
Communications and Networking, E. Hossain, Z. Han, and H. V. Poor,
Eds. Cambridge, U.K.: Cambridge Univ. Press, to be published.
[32] L. Zhao and A. Abur, Multiarea state estimation using synchronized
phasor measurements, IEEE Trans. Power Syst., vol. 20, no. 2, pp.
611617, May 2005.
Le Xie (S05M10) received the B.E. degree in elec-
trical engineering from Tsinghua University, Beijing,
China, in 2004, the M.Sc. degree in engineering sci-
ences from Harvard University, Cambridge, MA, in
June 2005, and the Ph.D. degree from the Electric
Energy Systems Group (EESG) in the Department
of Electrical and Computer Engineering at Carnegie
Mellon University, Pittsburgh, PA, in 2009.
He is an Assistant Professor in the Department of
Electrical and Computer Engineering at Texas A&M
University, College Station, where he is afliated
with the Electric Power and Power Electronics Group. His industry experience
includes an internship in 2006 at ISO-New England and an internship at Edison
Mission Energy Marketing and Trading in 2007. His research interests include
modeling, estimation and control of large-scale power systems, and electricity
markets.
Dae-Hyun Choi (S10) received the B.S. degree in
electrical engineering from Korea University, Seoul,
Korea, in 2002 and the M.Sc. degree in electrical and
computer engineering from Texas A&M University,
College Station, in 2008. He is pursuing the Ph.D.
degree in the Department of Electrical and Computer
Engineering at Texas A&M University.
From2002 to 2006, he was a researcher with Korea
Telecom (KT), Seoul, Korea, where he worked on
designing and implementing home network systems.
His research interests include power system state es-
timation, electricity markets, cyber-physical security of smart grid, and theory
and application of cyber-physical energy systems.
Soummya Kar (S05M10) received the B.Tech.
degree in electronics and electrical communication
engineering from the Indian Institute of Technology,
Kharagpur, India, in May 2005 and the Ph.D. degree
in electrical and computer engineering fromCarnegie
Mellon University, Pittsburgh, PA, in June 2010.
From June 2010 to May 2011, he was with the
Electrical Engineering Department at Princeton
University, Princeton, NJ, as a Postdoctoral Re-
search Associate. He is currently an Assistant
Research Professor of the Electrical and Computer
Engineering Department at Carnegie Mellon University. His research interests
include performance analysis and inference in large-scale networked systems,
adaptive stochastic systems, stochastic approximation, and large deviations.
H. Vincent Poor (S72M77SM82F87) re-
ceived the Ph.D. degree in electrical engineering
and computer science from Princeton University,
Princeton, NJ, in 1977.
From 1977 until 1990, he was on the faculty of the
University of Illinois at Urbana-Champaign. Since
1990, he has been on the faculty at Princeton, where
he is the Michael Henry Strater University Professor
of Electrical Engineering and Dean of the School of
Engineering and Applied Science. His research inter-
ests are in the areas of stochastic analysis, statistical
signal processing, and information theory, and their applications in wireless net-
works and related elds such as social networks and smart grid. Among his
publications in these areas are the recent books Classical, Semi-classical and
Quantum Noise (New York: Springer, 2012) and Smart Grid Communications
and Networking (Cambridge, U.K.: Cambridge Univ. Press, 2012).
Dr. Poor is a member of the National Academy of Engineering and the Na-
tional Academy of Sciences, a Fellow of the American Academy of Arts and
Sciences, and an International Fellow of the Royal Academy of Engineering
(U.K.). He is also a Fellowof the Institute of Mathematical Statistics, the Acous-
tical Society of America, and other organizations. In 1990, he served as Presi-
dent of the IEEE Information Theory Society, and in 20042007, he served as
the Editor-in-Chief of the IEEE TRANSACTIONS ON INFORMATION THEORY. He
received a Guggenheim Fellowship in 2002 and the IEEE Education Medal in
2005. Recent recognition of his work includes the 2010 IET Ambrose Fleming
Medal, the 2011 IEEE Eric E. Sumner Awards, the 2011 Society Award of the
IEEE Signal Processing Society, and honorary doctorates from the University
of Edinburgh and Aalborg University, conferred in 2011 and 2012, respectively.

Anda mungkin juga menyukai